Prediction of forest fires occurrences with area-level Poisson mixed models

https://doi.org/10.1016/j.jenvman.2015.02.009Get rights and content

Highlights

  • The methodology predicts the probability fire counts.

  • Bootstrap estimators are created for mean squared errors of fire predictions.

  • Bootstrap confidence intervals are used to calculate two-scenario fire variation.

  • The new tools study the effect of auxiliary variables in fire count reduction.

Abstract

The number of fires in forest areas of Galicia (north-west of Spain) during the summer period is quite high. Local authorities are interested in analyzing the factors that explain this phenomenon. Poisson regression models are good tools for describing and predicting the number of fires per forest areas. This work employs area-level Poisson mixed models for treating real data about fires in forest areas. A parametric bootstrap method is applied for estimating the mean squared errors of fires predictors. The developed methodology and software are applied to a real data set of fires in forest areas of Galicia.

Introduction

The number of human-caused bushfires has increased worldwide in the last years (Plucinski et al., 2013, Le Page et al., 2014, Krawchuk and Moritz, 2014). Policy makers and technicians request scientific models to explain the causality of fire and to establishing future scenarios of fire risk conditions, especially in one of the worst hit regions of the world, Mediterranean Europe (Rodrigues and de la Riva, 2014). This need has brought about the development of several prediction models (Martínez et al., 2009, Thompson and Calkin, 2011, Ager et al., 2014) which have focus on explaining spatio-temporal patterns that relate different variables (physiographic, infrastructural, socio-economic and weather) with ignition arson wildfires. The “number of fires” is a count variable, which is not continuous and the use of linear models is not appropriate. For this type of data, McCullagh and Nelder (1989) proposed an extension of linear models called generalized linear models (GLMs). GLMs assume that the distribution of the target variable belongs to the exponential distribution and therefore they can be used for Bernoulli, binomial and Poisson distributions, among others.

Poisson regressions and binomial-logit models are GLMs that are generally used for counts of rare events. This is the case of counting number of fires in a given territory and time period. Some papers introducing statistical models for counts of forest fires can be found in the literature. García et al. (1995) introduced a logit model to predict the number of fire-days in the Whitecourt Forest of Alberta. Mandallaz and Ye (1997) presented a general statistical methodology for the prediction of forest fires occurrences in the context of Poisson models. They applied their methodology to data from France, Italy, Portugal, and Switzerland. Wotton et al. (2003) developed Poisson regression predictive models for the daily number of fires in ecoregions of Ontario. Martínez et al., 2009, Chuvieco et al., 2010 and De Vicente and Crespo (2012) gave models for probability of occurrence of a fire.

The occurrence of the studied events might have some variability between areas that the GLMs cannot explain with the auxiliary variables. Generalized linear mixed models (GLMMs) allow to model the extra variability between areas by introducing a random effect. The estimation of GLMMs are computationally intensive and requires advanced numerical procedures. Some monographs about GLMMs are Demidenko, 2004, McCulloch et al., 2008 and Jiang (2007). Finney et al. (2009) analyzed the containment of large wild fires by using GLMMs. Díaz-Avalos et al. (2001) used GLMMs to study the effect of vegetation cover, elevation, slope, and precipitation on the probability of ignition in the Blue Mountains, Oregon, and to estimate the probability of ignition occurrence at different locations in space and in time.

This paper introduces an area-level Poisson mixed models for modeling the number of fires occurred in the forest areas of Galicia during the summer of 2007. The fitted GLMM is also used for predicting the number of fires under different scenario and therefore it gives a tool for decision making. The mean squared error (MSE) of the fire predictions was estimated by applying the parametric bootstrap method described in González-Manteiga et al., 2007, González-Manteiga et al., 2008. The proposed methodology for predicting fire counts and estimating prediction uncertainties is a new and useful contribution for forest engineers and policy makers.

The remainder of the paper is organized as follows. Section 2 introduces the area-level Poisson mixed model in a general context and the model-based predictors of number of fires in a general context. It also gives the bootstrap procedure for estimating the MSEs. Section 3 describes the situation and the data under study. The main idea is to look for the factors that explain the number of large fires in rural areas of Galicia. This section presents the statistical analysis and shows how the fitted model can be used for predicting number of fires under scenarios that are close to the one observed in Galicia during the summer of 2007. Section 4 gives some conclusions. The Appendix gives some mathematical details about the Laplace approximation fitting algorithm.

Section snippets

The model

This section introduces the Poisson mixed model employed in the data analysis of the case study. First we introduce some notation and assumptions. Let us assume that the region under study can be partitioned into D forest areas (domains). Let {vd:d=1,,D} be a set of random effects that are i.i.d. N(0,1). The distribution of the target variable yd (number of fires in the forest area d), conditioned to the random effect vd, isyd|vdPoiss(μd),d=1,,D,where μd>0 is the mean of the Poisson

Case study

The region of Galicia is in the northwest of the Spain (see Fig. 1). Galicia has a population of 2,795,422 (5.9% of the Spanish population) and a surface area of 29,574 km2 (5.8% of Spain). Galicia has the sixth greatest absolute forest areas among Autonomous Communities, with more than 1,424,094 ha of forest wooded representing 51% of Galician land cover and 7.7% of forest cover in Spain. Ratio of the forest surface is notably higher than the national average (55.2%) (MAGRAMA, 2014). Regarding

Discussion and conclusions

In recent years, several authors have studied the situation of wildfires in Galicia from an environmental perspective (Vega et al., 2011, Lombao et al., 2014), their financial consequences (Barreal et al., 2011) and the causality in conflictive areas (Fuentes-Santos et al., 2013) or in the whole region (Chas-Amil et al., 2015). There are also different studies on the causes and consequences of extraordinary situations, such as the case of the first half of August, 2006 (González-Alonso and

References (51)

  • M.P. Thompson et al.

    Uncertainty and risk in wildland fire management: a review

    J. Environ. Manag.

    (2011)
  • J. Vega et al.

    Pinus pinaster Ait. tree mortality following wildfire in Spain

    For. Ecol. Manag.

    (2011)
  • A.A. Ager et al.

    Wildfire risk estimation in the Mediterranean area

    Environmetrics

    (2014)
  • S. Bajocco et al.

    Evidence of selective burning in Sardinia (Italy): which land-cover classes do wildfires prefer?

    Landsc. Ecol.

    (2008)
  • J. Balsa-Barreiro et al.

    Socio-geographic analysis of wildland fires: causes of the 2006's wildfires in Galicia (Spain)

    For. Syst.

    (2013)
  • J. Barreal et al.

    Estudo da incidencia dos incendios en Galicia: Unha perspectiva socioeconomica

    Rev. Galega Econ.

    (2011)
  • X. Bruña García et al.

    Public participation: a need of forest planning

    IForest

    (2014)
  • A. Davison et al.

    Bootstrap Methods and Their Applications

    (2007)
  • J. De Vicente et al.

    A new wildland fire danger index for a Mediterranean region and some validation aspects

    Int. J. Wildland Fire

    (2012)
  • E. Demidenko

    Mixed Models: Theory and Applications

    (2004)
  • C. Díaz-Avalos et al.

    Spacetime modelling of lightning-caused ignitions in the Blue Mountains, Oregon

    Can. J. For. Res.

    (2001)
  • R. Díaz-Delgado et al.

    Spatial patterns of fire occurrence in Catalonia, NE Spain

    Landsc. Ecol.

    (2004)
  • M. Finney et al.

    Modeling containment of large wildfires using generalized linear mixed-model analysis

    For. Sci.

    (2009)
  • C. García et al.

    A logit model for predicting the daily occurrence of human caused forest-fires

    Int. J. Wildland Fire

    (1995)
  • I. Gómez-Vázquez et al.

    Conflicts as enhancers or barriers to the management of privately owned common land: a method to analyze the role of conflicts on a regional basis

    For. Policy Econ.

    (2009)
  • Cited by (42)

    • The importance of ground-truth and crowdsourcing data for the statistical and spatial analyses of the NASA FIRMS active fires in the Mediterranean Turkish forests

      2020, Remote Sensing Applications: Society and Environment
      Citation Excerpt :

      In addition, there is an increasing tendency in the number of fires in Europe, especially in the Mediterranean region. In Europe, more than 1600 wildfires were recorded in 2019; this is more than three times the average in the last decade (Euronews, 2019; Boubeta et al., 2015). As being located in the Mediterranean region, the number of annual forest fires in Turkey is also very high (Fig. 1) (Gonçalves et al., 2017; Turkish General Directory of Forestry, 2019).

    • Studying the occurrence and burnt area of wildfires using zero-one-inflated structured additive beta regression

      2018, Environmental Modelling and Software
      Citation Excerpt :

      Subsequently, Fuentes-Santos et al. (2015) used smooth bootstrap kernel intensity estimation for inhomogeneous spatial point processes to analyze the entire region in 2006, and concluded that the spatial distribution of arson and natural wildfires was different. Boubeta et al. (2015) applied Poisson mixed models with areas included as random effects to 63 forest areas constituting the basic structure of the fight against wildfires in Galicia, during the summer of 2007. The results showed significant differences between forest areas across the period of analysis, similar to our results.

    • Improving fire season definition by optimized temporal modelling of daily human-caused ignitions

      2018, Journal of Environmental Management
      Citation Excerpt :

      Our database contained 158,265 municipality-day observations, 77,097 in the Atlantic zone and 81,168 in the Mediterranean region. The spatial variables presented in Table 1 had been used in previous studies of HCFs occurrence modelling (i.e. Boubeta et al., 2015; Gonzalez-Olabarria et al., 2011; Martínez et al., 2009; Mhawej et al., 2015; Serra et al., 2014). The database included anthropic predictors, physiographic- and land use-related factors.

    • Unit-level and area-level small area estimation under heteroscedasticity using digital aerial photogrammetry data

      2018, Remote Sensing of Environment
      Citation Excerpt :

      Inference in remote sensing-supported forest inventories, i.e. the estimation of a population parameter and the associated uncertainty, using unit-level estimators has recently received more attention (Mandallaz, 2013; McRoberts et al., 2014; Saarela et al., 2015; Saarela et al., 2016; Gregoire et al., 2016; Chen et al., 2016; Mauro et al., 2016). Also, area-level estimators are becoming more studied (Goerndt et al., 2011, 2013; Boubeta et al., 2015; Magnussen et al., 2017). However, unit- and area-level estimators are seldom compared (Hidiroglou and You, 2016).

    View all citing articles on Scopus

    Supported by the grants MTM2013-41383-P, MTM2012-37077-C02-01, MTM2011-22392 and MTM2008-03010 of the Spanish “Ministerio de Ciencia e Innovación”, by “Xunta de Galicia” CN2012/130 and 07MRU035291PR, by “Ministerio del Medio Ambiente, Rural y Marino” PSE-310000-2009-4 and by COST Action/UE COST-OC-2008-1-2124.

    View full text